twitter storm源码走读（二）
topology提交过程分析 概要 storm cluster可以想像成为一个工厂，nimbus主要负责从外部接收订单和任务分配。除了从外部接单，nimbus还要将这些外部订单转换成为内部工作分配，这个时候nimbus充当了调度室的角色。supervisor作为中层干部，职责就是生产车间的主任，他的日常工作就是时刻等待着调度到给他下达新的工作。作为车间主任，supervisor领到的活是不用自己亲力亲为去作的，他手下有着一班的普通工人。supervisor对这些工人只会喊两句话，开工，收工。注意，讲收工的时候并不意味着worker手上的活已经干完了，只是进入休息状态而已。 topology的提交过程涉及到以下角色。 storm client 负责将用户创建的topology提交到nimbus nimbus 通过thrift接口接收用户提交的topology supervisor根据zk接口上提示的消息下载最新的任务安排，并负责启动worker workerworker内可以运行task,这些task要么属于bolt类型，要么属于spout类型 executorexecutor是一个个运行的线程，同一个executor内可以运行同一种类型的task,即一个线程中的task要么全部是bolt类型，要么全部是spout类型 一个worker等同于一个进程，一个executor等同于一个线程，同一个线程中能够运行一或多个tasks。在0.8.0版之前，一个task是对应于一个线程的，在0.8.0版本中引入了executor概念，变化引入之后，task与thread之间的一一对应关系就取消了，同时在zookeeper server中原本存在的tasks-subtree也消失了，有关这个变化，可以参考http://storm-project.net/2012/08/02/storm080-released.html storm client storm client需要执行下面这句指令将要提交的topology提交给storm cluster 假设jar文件名为storm-starter-0.0.1-snapshot-standalone.jar,启动程序为 storm.starter.ExclamationTopology,给这个topology起的名称为exclamationTopology. #./storm jar $HOME/working/storm-starter/target/storm-starter-0.0.1-SNAPSHOT-standalone.jar storm.starter.ExclamationTopology exclamationTopology 这么短短的一句话对于storm client来说，究竟意味着什么呢？ 源码面前是没有任何秘密可言的，那好打开storm client的源码文件 def jar(jarfile, klass, *args): """Syntax: [storm jar topology-jar-path class ...] Runs the main method of class with the specified arguments. The storm jars and configs in ~/.storm are put on the classpath. The process is configured so that StormSubmitter (http://nathanmarz.github.com/storm/doc/backtype/storm/StormSubmitter.html) will upload the jar at topology-jar-path when the topology is submitted. """ exec_storm_class( klass, jvmtype="-client", extrajars=[jarfile, USER_CONF_DIR, STORM_DIR + "/bin"], args=args, jvmopts=["-Dstorm.jar=" + jarfile]) def exec_storm_class(klass, jvmtype="-server", jvmopts=[], extrajars=[], args=[], fork=False): global CONFFILE all_args = [ "java", jvmtype, get_config_opts(), "-Dstorm.home=" + STORM_DIR, "-Djava.library.path=" + confvalue("java.library.path", extrajars), "-Dstorm.conf.file=" + CONFFILE, "-cp", get_classpath(extrajars), ] + jvmopts + [klass] + list(args) print "Running: " + " ".join(all_args) if fork: os.spawnvp(os.P_WAIT, "java", all_args) else: os.execvp("java", all_args) # replaces the current process and never returns exec_storm_class说白了就是要运行传进来了的WordCountTopology类中main函数，再看看main函数的实现 public static void main(String[] args) throws Exception { TopologyBuilder builder = new TopologyBuilder(); builder.setSpout("spout", new RandomSentenceSpout(), 5); builder.setBolt("split", new SplitSentence(), 8).shuffleGrouping("spout"); builder.setBolt("count", new WordCount(), 12).fieldsGrouping("split", new Fields("word")); Config conf = new Config(); conf.setDebug(true); if (args != null && args.length > 0) { conf.setNumWorkers(3); StormSubmitter.submitTopology(args[0], conf, builder.createTopology()); } } 对于storm client侧来说，最主要的函数StormSubmitter露出了真面目，submitTopology才是我们真正要研究的重点。 public static void submitTopology(String name, Map stormConf, StormTopology topology, SubmitOptions opts) throws AlreadyAliveException, InvalidTopologyException { if(!Utils.isValidConf(stormConf)) { throw new IllegalArgumentException("Storm conf is not valid. Must be json-serializable"); } stormConf = new HashMap(stormConf); stormConf.putAll(Utils.readCommandLineOpts()); Map conf = Utils.readStormConfig(); conf.putAll(stormConf); try { String serConf = JSONValue.toJSONString(stormConf); if(localNimbus!=null) { LOG.info("Submitting topology " + name + " in local mode"); localNimbus.submitTopology(name, null, serConf, topology); } else { NimbusClient client = NimbusClient.getConfiguredClient(conf); if(topologyNameExists(conf, name)) { throw new RuntimeException("Topology with name `" + name + "` already exists on cluster"); } submitJar(conf); try { LOG.info("Submitting topology " + name + " in distributed mode with conf " + serConf); if(opts!=null) { client.getClient().submitTopologyWithOpts(name, submittedJar, serConf, topology, opts); } else { // this is for backwards compatibility client.getClient().submitTopology(name, submittedJar, serConf, topology); } } catch(InvalidTopologyException e) { LOG.warn("Topology submission exception", e); throw e; } catch(AlreadyAliveException e) { LOG.warn("Topology already alive exception", e); throw e; } finally { client.close(); } } LOG.info("Finished submitting topology: " + name); } catch(TException e) { throw new RuntimeException(e); } } submitTopology函数其实主要就干两件事，一上传jar文件到storm cluster，另一件事通知storm cluster文件已经上传完毕，你可以执行某某某topology了. 先看上传jar文件对应的函数submitJar,其调用关系如下图所示 再看第二步中的调用关系，图是我用tikz/pgf写的，生成的是pdf格式。 在上述两幅调用关系图中，处于子树位置的函数都曾在storm.thrift中声明，如果此刻已经忘记了的点话，可以翻看一下前面1.3节中有关storm.thrift的描述。client侧的这些函数都是由thrift自动生成的。 由于篇幅和时间的关系，在storm client侧submit topology的时候，非常重要的函数还有TopologyBuilder.java中的源码。 nimbus storm client侧通过thrift接口向nimbus发送了了jar并且通过预先定义好的submitTopologyWithOpts来处理上传的topology，那么nimbus是如何一步步的进行文件接收并将其任务细化最终下达给supervisor的呢。 submitTopologyWithOpts 一切还是要从thrift说起，supervisor.clj中的service-handler具体实现了thrift定义的Nimbus接口，代码这里就不罗列了，太占篇幅。主要看其是如何实现submitTopologyWithOpts (^void submitTopologyWithOpts [this ^String storm-name ^String uploadedJarLocation ^String serializedConf ^StormTopology topology ^SubmitOptions submitOptions] (try (assert (not-nil? submitOptions)) (validate-topology-name! storm-name) (check-storm-active! nimbus storm-name false) (.validate ^backtype.storm.nimbus.ITopologyValidator (:validator nimbus) storm-name (from-json serializedConf) topology) (swap! (:submitted-count nimbus) inc) (let [storm-id (str storm-name "-" @(:submitted-count nimbus) "-" (current-time-secs)) storm-conf (normalize-conf conf (-> serializedConf from-json (assoc STORM-ID storm-id) (assoc TOPOLOGY-NAME storm-name)) topology) total-storm-conf (merge conf storm-conf) topology (normalize-topology total-storm-conf topology) topology (if (total-storm-conf TOPOLOGY-OPTIMIZE) (optimize-topology topology) topology) storm-cluster-state (:storm-cluster-state nimbus)] (system-topology! total-storm-conf topology) ;; this validates the structure of the topology (log-message "Received topology submission for " storm-name " with conf " storm-conf) ;; lock protects against multiple topologies being submitted at once and ;; cleanup thread killing topology in b/w assignment and starting the topology (locking (:submit-lock nimbus) (setup-storm-code conf storm-id uploadedJarLocation storm-conf topology) (.setup-heartbeats! storm-cluster-state storm-id) (let [thrift-status->kw-status {TopologyInitialStatus/INACTIVE :inactive TopologyInitialStatus/ACTIVE :active}] (start-storm nimbus storm-name storm-id (thrift-status->kw-status (.get_initial_status submitOptions)))) (mk-assignments nimbus))) (catch Throwable e (log-warn-error e "Topology submission exception. (topology name='" storm-name "')") (throw e)))) storm cluster在zookeeper server上创建的目录结构。目录结构相关的源文件是config.clj. 白话一下上面这个函数的执行逻辑，对上传的topology作必要的检测，包括名字，文件内容及格式，好比你进一家公司上班之前做的体检。这些工作都完成之后进入关键区域，是进入关键区域所以上锁，呵呵。 normalize-topology (defn all-components [^StormTopology topology] (apply merge {} (for [f thrift/STORM-TOPOLOGY-FIELDS] (.getFieldValue topology f) ))) 一旦列出所有的components,就可以读出这些component的配置信息。 mk-assignments 在这关键区域内执行的重点就是函数mk-assignments，mk-assignment有两个主要任务，第一是计算出有多少task,即有多少个spout,多少个bolt，第二就是在刚才的计算基础上通过调用zookeeper应用接口，写入assignment，以便supervisor感知到有新的任务需要认领。 先说第二点，因为逻辑简单。在mk-assignment中执行如下代码在zookeeper中设定相应的数据以便supervisor能够感知到有新的任务产生 (doseq [[topology-id assignment] new-assignments :let [existing-assignment (get existing-assignments topology-id) topology-details (.getById topologies topology-id)]] (if (= existing-assignment assignment) (log-debug "Assignment for " topology-id " hasn't changed") (do (log-message "Setting new assignment for topology id " topology-id ": " (pr-str assignment)) (.set-assignment! storm-cluster-state topology-id assignment) ))) 调用关系如下图所示 而第一点涉及到的计算相对繁杂，需要一一仔细道来。其实第一点中非常重要的课题就是如何进行任务的分发，即scheduling. 也许你已经注意到目录src/clj/backtype/storm/scheduler，或者注意到storm.yaml中与scheduler相关的配置项。那么这个scheduler到底是在什么时候起作用的呢。mk-assignments会间接调用到这么一个名字看起来奇怪异常的函数。compute-new-topology->executor->node+por，也就是在这么很奇怪的函数内，scheduler被调用 _ (.schedule (:scheduler nimbus) topologies cluster) new-scheduler-assignments (.getAssignments cluster) ;; add more information to convert SchedulerAssignment to Assignment new-topology->executor->node+port (compute-topology->executor->node+port new-scheduler-assignments)] schedule计算出来的assignments保存于Cluster.java中，这也是为什么new-scheduler-assignment要从其中读取数据的缘由所在。有了assignment，就可以计算出相应的node和port，其实就是这个任务应该交由哪个supervisor上的worker来执行。 storm在zookeeper server上创建的目录结构如下图所示 有了这个目录结构，现在要解答的问题是在topology在提交的时候要写哪几个目录？assignments目录下会新创建一个新提交的topology的目录，在这个topology中需要写的数据，其数据结构是什么样子？ supervisor 一旦有新的assignment被写入到zookeeper中，supervisor中的回调函数mk-synchronize-supervisor立马被唤醒执行 主要执行逻辑就是读入zookeeper server中新的assignments全集与已经运行与本机上的assignments作比较，区别出哪些是新增的。在sync-processes函数中将运行具体task的worker拉起。 要想讲清楚topology提交过程中，supervisor需要做哪些动作，最主要的是去理解下面两个函数的处理逻辑。 mk-synchronize-supervisor当在zookeeper server的assignments子目录内容有所变化时，supervisor收到相应的notification, 处理这个notification的回调函数即为mk-synchronize-supervisor，mk-sychronize-supervisor读取所有的assignments即便它不是由自己处理，并将所有assignment的具体信息读出。尔后判断分析出哪些assignment是分配给自己处理的，在这些分配的assignment中，哪些是新增的。知道了新增的assignment之后，从nimbus的相应目录下载jar文件，用户自己的处理逻辑代码并没有上传到zookeeper server而是在nimbus所在的机器硬盘上。 sync-processesmk-synchronize-supervisor预处理过完与assignment相关的操作后，将真正启动worker的动作交给event-manager, event-manager运行在另一个独立的线程中，这个线程中进行处理的一个主要函数即sync-processes. sync-processes会将当前运行着的worker全部kill,然后指定新的运行参数，重新拉起worker. (defn mk-synchronize-supervisor [supervisor sync-processes event-manager processes-event-manager] (fn this [] (let [conf (:conf supervisor) storm-cluster-state (:storm-cluster-state supervisor) ^ISupervisor isupervisor (:isupervisor supervisor) ^LocalState local-state (:local-state supervisor) sync-callback (fn [& ignored] (.add event-manager this)) assignments-snapshot (assignments-snapshot storm-cluster-state sync-callback) storm-code-map (read-storm-code-locations assignments-snapshot) downloaded-storm-ids (set (read-downloaded-storm-ids conf)) ;;read assignments from zookeeper all-assignment (read-assignments assignments-snapshot (:assignment-id supervisor)) new-assignment (->> all-assignment (filter-key #(.confirmAssigned isupervisor %))) ;;task在assignment中 assigned-storm-ids (assigned-storm-ids-from-port-assignments new-assignment) existing-assignment (.get local-state LS-LOCAL-ASSIGNMENTS)] (log-debug "Synchronizing supervisor") (log-debug "Storm code map: " storm-code-map) (log-debug "Downloaded storm ids: " downloaded-storm-ids) (log-debug "All assignment: " all-assignment) (log-debug "New assignment: " new-assignment) ;; download code first ;; This might take awhile ;; - should this be done separately from usual monitoring? ;; should we only download when topology is assigned to this supervisor? (doseq [[storm-id master-code-dir] storm-code-map] (when (and (not (downloaded-storm-ids storm-id)) (assigned-storm-ids storm-id)) (log-message "Downloading code for storm id " storm-id " from " master-code-
